Wearable Biomarker Discovery, With a Human in the Loop
— by Vivax
Google's August 2026 Biomarker Discovery Framework, called CoDaS in its associated paper, uses a human-supervised, multi-agent loop to prioritize candidate…
Google's August 2026 Biomarker Discovery Framework, called CoDaS in its associated paper, uses a human-supervised, multi-agent loop to prioritize candidate digital biomarkers from wearable data. The goal is discovery: narrow a large hypothesis space before expensive prospective validation.
Across three cohorts totaling 9,279 participant-observations, the framework reports 41 mental-health and 25 metabolic candidates. Those counts are candidates, not diagnoses or cleared biomarkers. Wearable signals are correlated with behavior, medication, device use and missingness, so a plausible association can be badly confounded.
The workflow combines literature-grounded proposal, statistical testing and adversarial review. Agents propose features, inspect robustness and challenge leakage; human researchers supervise the loop and decide which hypotheses deserve follow-up. The multi-agent design is a way to organize research, not a replacement for clinical judgment.
Credibility depends on controls: participant-level splits, time-aware validation, adjustment for multiple testing, negative controls and checks that a device or cohort artifact is not carrying the signal. Even a clean retrospective association can fail when sensors, routines or population prevalence change.
The crucial sentence is what the framework does not claim. A candidate still needs independent cohorts, preregistered analysis, prospective collection, clinical utility testing and a clearly defined intervention or decision. Discovery can accelerate science while leaving clinical validity completely open.
Vivax sees a close fit with a grounded clinical world model: sensor streams belong in a longitudinal context and their provenance must remain visible. We would use this kind of system to prioritize questions, not to turn a wearable correlation into a diagnosis or treatment recommendation.